Spoken Arabic Regional Archive (SARA)

SARA is a database, which consists of a set of spontaneous not pre-specified colloquial phrases in everyday life and life situations that are collected from media shows, episodes and films published on YouTube played by native speakers with three different Arabic dialects and accents. SARA's dialects are the Egyptian dialect (EGY), the Arabian Peninsula dialect (ARP) and the Levantine dialect (LEV). In addition, within each dialect group there are a number of different accents. Each utterance contains a one-speaker speech, the number of speakers within the dialect and the number of utterances by a speaker is unknown. This dataset can be used in speech, speaker, dialect and accent recognition applications. SARA dataset contains only adult speakers to avoid the improper pronunciation of the children that can affect the detection process. The dataset samples are variant in length from 3 to 7 seconds in order to verify the minimum time in which we can determine the speaker dialect or accent when the speaker speaks in free talk, which is the main scope of this research.SARA是一款语音数据库，收录了日常及真实生活场景中的自发、未预先设定的口语化短语与话语，其数据采集自YouTube平台发布的媒体节目、剧集及电影，由三种不同阿拉伯方言口音的母语使用者录制。SARA涵盖的方言包括埃及方言（Egyptian dialect, EGY）、阿拉伯半岛方言（Arabian Peninsula dialect, ARP）以及黎凡特方言（Levantine dialect, LEV）。此外，每个方言组内还包含多种不同的口音。每条语音片段均为单发言人发声，目前尚未明确各方言组内的总发言人数，以及每位发言人对应的语音片段数量。本数据集可应用于语音识别、发言人识别、方言识别及口音识别相关任务。SARA数据集仅收录成年发言人的语音数据，以规避儿童发音不规范对识别流程造成的负面影响。数据集的语音样本时长跨度为3至7秒，旨在验证在自由会话场景下，可用于准确判定发言人方言或口音的最短时长——这也是本研究的核心研究范畴。